Research Of Visual SLAM Based On Multi-Camera In The Dynamic Environment

Nowadays,Robotics,Virtual Reality,Augmented Reality are all very popular in both academic and industry,and they all need SLAM algorithm to achieve localization.SLAM algorithm based on 2D laser sensor has been highly developed in recent decades,while Visual-based SLAM still have many limitations,such as cannot achieve robust localization in dynamic environment.More researchers are focusing on visual SLAM nowadays.Those SLAM algorithms that use stereo camera system,are studied widely since stereo camera system not only is more robust than monocular camera system,but also could return the true deep distances through depth estimation algorithm.Therefore,it is of great significance for us to research visual SLAM algorithm with multi-camera system in dynamic environment.In this paper,we mainly research visual SLAM algorithm based on stereo camera system in dynamic environment,the main work is as follows:(1)Introduce the pinhole camera model and stereo camera model,which are commonly used in SLAM algorithms,and derived the coordinate system transformation from image coordinate system to the world coordinate system.Comparing the speed of extracting image features with different feature description operators,do experiment to extract ORB features of the image and the images of different frames are combined according to the feature matching algorithm.With the help of camera's intrinsic matrix and extrinsic matrix in stereo camera system,the direct method is used to estimate the depth of the features in the real world.Then,estimate the camera's pose changes between different frames by Pn P algorithm.(2)Using the method of machine learning to learn the vehicle characteristics of highway environment,a "two-step iteration" vehicle feature screening method is proposed to solve the problem of lack of features caused by erroneous recognition of dynamic objects in the environment.(3)Based on the classic visual SLAM framework,a new visual SLAM framework for dynamic scenes is proposed and validated in the highway environment.Using the KITTI data set for experiments,the positioning accuracy has been significantly improved,and the accurate positioning in dynamic environment and the construction of a relatively accurate environmental map have been achieved.